torchtune icon indicating copy to clipboard operation
torchtune copied to clipboard

Is there a way to get PreferenceDataset to support packing?

Open iqiancheng opened this issue 1 year ago • 1 comments

When I try to set packed=True, it's rejected by the parent class PreferenceDataset which raises a ValueError with "Packed is currently not supported for preference datasets."

Is there a way to get PreferenceDataset to support packing?

iqiancheng avatar Apr 23 '25 06:04 iqiancheng

Hi @iqiancheng, sample packing doesn't translate to preference datasets easily. There is prefix-sharing which I think looks interesting, but there no concrete plans to add it for now.

salmanmohammadi avatar Apr 23 '25 09:04 salmanmohammadi